DOCSLIB.ORG

Design of Observational Studies (Springer Series in Statistics)

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Design of Observational Studies (Springer Series in Statistics) Springer Series in Statistics Advisors: P. Bickel, P. Diggle, S. Fienberg, U. Gather, I. Olkin, S. Zeger For other titles published in this series, go to http://www.springer.com/series/692 Paul R. Rosenbaum Design of Observational Studies 123 Paul R. Rosenbaum Statistics Department Wharton School University of Pennsylvania Philadelphia, PA 19104-6340 USA [email protected] ISBN 978-1-4419-1212-1 e-ISBN 978-1-4419-1213-8 DOI 10.1007/978-1-4419-1213-8 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2009938109 c Springer Science+Business Media, LLC 2010 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper Springer is a part of Springer Science+Business Media (www.springer.com). For Judy “Simplicity of form is not necessarily simplicity of experience.” Robert Morris, writing about art. “Simplicity is not a given. It is an achievement.” William H. Gass, writing about literature. “Simplicity . is a very important matter which must be constantly borne in mind.” Sir David Cox, writing about experiments. Preface An observational study is an empiric investigation of effects caused by treatments when randomized experimentation is unethical or infeasible. The quality and strength of evidence provided by an observational study is determined largely by its design. Excellent methods of analysis will not salvage a poorly designed study. The line between design and analysis is easier to draw in practice than it is in the- ory. In practice, the design of an observational study consists of all activities that precede the examination or use of those outcome measures that will be the basis for the study’s conclusions. Unlike experiments, in some observational studies, the outcomes may exist as measurements prior to the design of the study; it is their ex- amination and use, not their existence, that separates design from analysis. Aspects of design include the framing of scientific questions to permit empirical investiga- tion, the choice of a context in which to conduct the investigation, decisions about what data to collect, where and how to collect it, matching to remove bias from measured covariates, strategies and tactics to limit uncertainty caused by covariates not measured, and sample splitting to guide design using individuals who will not be included in the final analysis. In practice, design ends and analysis begins when outcomes are examined for individuals who will be the basis of the study’s conclu- sions. An observational study that begins by examining outcomes is a formless, undisciplined investigation that lacks design. In theory, design anticipates analysis. Analysis is ever present in design, as any goal is ever present in any organized effort, as a goal is necessary to organize effort. One seeks to ask questions and collect data so that results will be decisive when analyzed. To end well, how should we begin? Philadelphia, PA Paul Rosenbaum 5 August 2009 ix Acknowledgments I am in debt to many people: to Jeff Silber, Dylan Small, and Ruth Heller for re- cent collaborations I describe here in detail; to my teacher, adviser, and coauthor Don Rubin, from whom I learned a great deal; to Ben Hansen, Bo Lu and Robert Greevy for making network optimization algorithms for matching generally accessi- ble inside R; to colleagues, coauthors or former students Katrina Armstrong, Susan Bakewell-Sachs, Lisa Bellini, T. Behringer, Avital Cnaan, Shoshana Daniel, Gabriel Escobar, Orit Even-Shoshan, Joe Gastwirth, Robert Greevy, Sam Gu, Amelia Hav- iland, Robert Hornik, Abba Krieger, Marshall Joffe, Yunfei Paul Li, Scott Lorch, Bo Lu, Barbara Medoff-Cooper, Lanyu Mi, Andrea Millman, Kewei Ming, Dan Nagin, Dan Polsky, Kate Propert, Tom Randall, Amy Rosen, Richard Ross, Sandy Schwartz, Tom Ten Have, Richard Tremblay, Kevin Volpp,Yanli Wang, Frank Yoon, and Elaine Zanutto, for collaborations I describe more briefly; to Judith McDonald, John Newell, Luke Keele, Dylan Small, and anonymous reviewers for comments on drafts of the book; to Joshua Angrist, David Card, Susan Dynarski, Alan Krueger, and Victor Lavy for making micro-data from their research available in one form or another. Parts of this book were written while I was on sabbatical. The hospitality of the Department of Statistics at Columbia University and the Department of Eco- nomics and the Statistics Program at the National University of Ireland at Galway are gratefully acknowledged. The work was supported in part by the Methodology, Measurement and Statistics Program of the U.S. National Science Foundation. Of course, my greatest debts are to Judy, Sarah, Hannah, and Aaron. xi Contents Part I Beginnings 1 Dilemmas and Craftsmanship ................................... 3 1.1 Those Confounded Vitamins ................................. 3 1.2 Cochran’s Basic Advice ..................................... 4 1.2.1 Treatments, covariates, outcomes ....................... 5 1.2.2 How were treatments assigned? . ....................... 5 1.2.3 Were treated and control groups comparable? . .......... 5 1.2.4 Eliminating plausible alternatives to treatment effects . ..... 6 1.2.5 Exclusion criteria . ................................. 6 1.2.6 Exiting a treatment group after treatment assignment . ..... 7 1.2.7 Study protocol. ..................................... 7 1.3 Maimonides’ Rule .......................................... 7 1.4 Seat Belts in Car Crashes . ................................. 9 1.5 Money for College ......................................... 10 1.6 Nature’s ‘Natural Experiment’ ................................ 11 1.7 What This Book Is About . ................................. 13 1.8 Further Reading . .......................................... 18 References . ................................................... 18 2 Causal Inference in Randomized Experiments ..................... 21 2.1 Two Versions of the National Supported Work Experiment . ..... 21 2.1.1 A version with 185 pairs and a version with 5 pairs . ..... 21 2.1.2 Basic notation . ..................................... 23 2.2 Treatment Effects in Randomized Experiments .................. 25 2.2.1 Potential responses under alternative treatments . .......... 25 2.2.2 Covariates and outcomes . ............................ 26 2.2.3 Possible treatment assignments and randomization . ..... 27 2.2.4 Interference between units . ............................ 28 2.3 Testing the Null Hypothesis of No Treatment Effect .............. 29 2.3.1 Treated−control differences when the null hypothesis is true 29 xiii xiv Contents 2.3.2 The randomization distribution of the mean difference ..... 31 2.3.3 The randomization distribution of Wilcoxon’s statistic ..... 36 2.4 Testing Other Hypotheses; Confidence Intervals; Point Estimates. 40 2.4.1 Testing a constant, additive treatment effect .............. 40 2.4.2 Confidence intervals for a constant, additive effect . ..... 41 2.4.3 Hodges-Lehmann point estimates of effect . .............. 43 2.4.4 Testing general hypotheses about treatment effects . ..... 44 2.4.5 Multiplicative effects; Tobit effects . ................... 46 2.5 Attributable Effects ......................................... 49 2.6 Internal and External Validity ................................ 56 2.7 Summary . .............................................. 57 2.8 Further Reading . .......................................... 57 2.9 Appendix: Randomization Distribution of m-statistics . .......... 58 References . ................................................... 61 3 Two Simple Models for Observational Studies ..................... 65 3.1 The Population Before Matching . ............................ 65 3.2 The Ideal Matching ......................................... 66 3.3 A Na¨ıve Model: People Who Look Comparable Are Comparable . 70 3.4 Sensitivity Analysis: People Who Look Comparable May Differ . 76 3.5 Welding Fumes and DNA Damage ............................ 79 3.6 Bias Due to Incomplete Matching . ............................ 85 3.7 Summary . .............................................. 86 3.8 Further Reading . .......................................... 87 3.9 Appendix: Exact Computations for Sensitivity Analysis .......... 88 References . ................................................... 90 4 Competing Theories Structure Design ............................ 95 4.1 How Stones Fall . .......................................... 95 4.2 The Permanent-Debt Hypothesis . ............................ 98 4.3 Guns and Misdemeanors . .................................100 4.4 The Dutch Famine of 1944–1945 . ............................100 4.5 Replicating Effects and Biases ................................101 4.6 Reasons for Effects .........................................104 4.7 The Drive for System . .....................................108 4.8 Further Reading . ..........................................109
Load more

Recommended publications
  • 19 Apr 2021 Model Combinations Through Revised Base-Rates
    Model combinations through revised base-rates Fotios Petropoulos School of Management, University of Bath UK and Evangelos Spiliotis Forecasting and Strategy Unit, School of Electrical and Computer Engineering, National Technical University of Athens, Greece and Anastasios Panagiotelis Discipline of Business Analytics, University of Sydney, Australia April 21, 2021 Abstract Standard selection criteria for forecasting models focus on information that is calculated for each series independently, disregarding the general tendencies and performances of the candidate models. In this paper, we propose a new way to sta- tistical model selection and model combination that incorporates the base-rates of the candidate forecasting models, which are then revised so that the per-series infor- arXiv:2103.16157v2 [stat.ME] 19 Apr 2021 mation is taken into account. We examine two schemes that are based on the pre- cision and sensitivity information from the contingency table of the base rates. We apply our approach on pools of exponential smoothing models and a large number of real time series and we show that our schemes work better than standard statisti- cal benchmarks. We discuss the connection of our approach to other cross-learning approaches and offer insights regarding implications for theory and practice. Keywords: forecasting, model selection/averaging, information criteria, exponential smooth- ing, cross-learning. 1 1 Introduction Model selection and combination have long been fundamental ideas in forecasting for business and economics (see Inoue and Kilian, 2006; Timmermann, 2006, and ref- erences therein for model selection and combination respectively). In both research and practice, selection and/or the combination weight of a forecasting model are case-specific.
    [Show full text]
  • Exponential Smoothing – Trend & Seasonal
    NCSS Statistical Software NCSS.com Chapter 467 Exponential Smoothing – Trend & Seasonal Introduction This module forecasts seasonal series with upward or downward trends using the Holt-Winters exponential smoothing algorithm. Two seasonal adjustment techniques are available: additive and multiplicative. Additive Seasonality Given observations X1 , X 2 ,, X t of a time series, the Holt-Winters additive seasonality algorithm computes an evolving trend equation with a seasonal adjustment that is additive. Additive means that the amount of the adjustment is constant for all levels (average value) of the series. The forecasting algorithm makes use of the following formulas: at = α( X t − Ft−s ) + (1 − α)(at−1 + bt−1 ) bt = β(at − at−1 ) + (1 − β)bt−1 Ft = γ ( X t − at ) + (1 − γ )Ft−s Here α , β , and γ are smoothing constants which are between zero and one. Again, at gives the y-intercept (or level) at time t, while bt is the slope at time t. The letter s represents the number of periods per year, so the quarterly data is represented by s = 4 and monthly data is represented by s = 12. The forecast at time T for the value at time T+k is aT + bT k + F[(T +k−1)/s]+1 . Here [(T+k-1)/s] is means the remainder after dividing T+k-1 by s. That is, this function gives the season (month or quarter) that the observation came from. Multiplicative Seasonality Given observations X1 , X 2 ,, X t of a time series, the Holt-Winters multiplicative seasonality algorithm computes an evolving trend equation with a seasonal adjustment that is multiplicative.
    [Show full text]
  • Exponential Smoothing Statlet.Pdf
    STATGRAPHICS Centurion – Rev. 11/21/2013 Exponential Smoothing Statlet Summary ......................................................................................................................................... 1 Data Input........................................................................................................................................ 2 Statlet .............................................................................................................................................. 3 Output ............................................................................................................................................. 4 Calculations..................................................................................................................................... 6 Summary This Statlet applies various types of exponential smoothers to a time series. It generates forecasts with associated forecast limits. Using the Statlet controls, the user may interactively change the values of the smoothing parameters to examine their effect on the forecasts. The types of exponential smoothers included are: 1. Brown’s simple exponential smoothing with smoothing parameter . 2. Brown’s linear exponential smoothing with smoothing parameter . 3. Holt’s linear exponential smoothing with smoothing parameters and . 4. Brown’s quadratic exponential smoothing with smoothing parameter . 5. Winters’ seasonal exponential smoothing with smoothing parameters and . Sample StatFolio: expsmoothing.sgp Sample Data The data
    [Show full text]
  • LECTURE 2 MOVING AVERAGES and EXPONENTIAL SMOOTHING OVERVIEW This Lecture Introduces Time-Series Smoothing Forecasting Methods
    Business Conditions & Forecasting – Exponential Smoothing Dr. Thomas C. Chiang LECTURE 2 MOVING AVERAGES AND EXPONENTIAL SMOOTHING OVERVIEW This lecture introduces time-series smoothing forecasting methods. Various models are discussed, including methods applicable to nonstationary and seasonal time-series data. These models are viewed as classical time-series model; all of them are univariate. LEARNING OBJECTIVES • Moving averages • Forecasting using exponential smoothing • Accounting for data trend using Holt's smoothing • Accounting for data seasonality using Winter's smoothing • Adaptive-response-rate single exponential smoothing 1. Forecasting with Moving Averages The naive method discussed in Lecture 1 uses the most recent observations to forecast future ˆ values. That is, Yt+1 = Yt. Since the outcomes of Yt are subject to variations, using the mean value is considered an alternative method of forecasting. In order to keep forecasts updated, a simple moving-average method has been widely used. 1.1. The Model Moving averages are developed based on an average of weighted observations, which tends to smooth out short-term irregularity in the data series. They are useful if the data series remains fairly steady over time. Notations ˆ M t ≡ Yt+1 - Moving average at time t , which is the forecast value at time t+1, Yt - Observation at time t, ˆ et = Yt − Yt - Forecast error. A moving average is obtained by calculating the mean for a specified set of values and then using it to forecast the next period. That is, M t = (Yt + Yt−1 + ⋅⋅⋅ + Yt−n+1 ) n (1.1.1) M t−1 = (Yt−1 +Yt−2 + ⋅⋅⋅+Yt−n ) n (1.1.2) Business Conditions & Forecasting Dr.
    [Show full text]
  • A Simple Method to Adjust Exponential Smoothing Forecasts for Trend and Seasonality
    Southern Methodist University SMU Scholar Historical Working Papers Cox School of Business 1-1-1993 A Simple Method to Adjust Exponential Smoothing Forecasts for Trend and Seasonality Marion G. Sobol Southern Methodist University Jim Collins Southern Methodist University Follow this and additional works at: https://scholar.smu.edu/business_workingpapers Part of the Business Commons This document is brought to you for free and open access by the Cox School of Business at SMU Scholar. It has been accepted for inclusion in Historical Working Papers by an authorized administrator of SMU Scholar. For more information, please visit http://digitalrepository.smu.edu. A SIMPLE METHOD TO ADJUST EXPONENTIAL SMOOTHING FORECASTS FOR TREND AND SEASONALITY Working Paper 93-0801* by Marion G. Sobol Jim Collins Marion G. Sobol Jim Collins Edwin L. Cox School of Business Southern Methodist University Dallas, Texas 75275 * This paper represents a draft of work in progress by the authors and is being sent to you for information and review. Responsibility for the contents rests solely with the authors and may not be reproduced or distributed without their written consent. Please address all correspondence to Marion G. Sobol. Abstract Early exponential smoothing techniques did not adjust for seasonality and trend. Later Holt and Winters developed methods to include these factors but they require the arbitrary choice of three smoothing constants '( , ~ and t. Our technique requires only alpha and an estimate of trend and seasonality as done for decomposition analysis. Since seasonality and trend are relatively constant why keep reestimating them - just include them. Moreover this technique is easy to perfonn on a standard spreadsheet.
    [Show full text]
  • 3. Regression & Exponential Smoothing
    3. Regression & Exponential Smoothing 3.1 Forecasting a Single Time Series Two main approaches are traditionally used to model a single time series z1, z2, . , zn 1. Models the observation zt as a function of time as zt = f(t, β) + εt where f(t, β) is a function of time t and unknown coefficients β, and εt are uncorrelated errors. 1 ∗ Examples: – The constant mean model: zt = β + εt – The linear trend model: zt = β0 + β1t + εt – Trigonometric models for seasonal time series 2π 2π z = β + β sin t + β cos t + ε t 0 1 12 2 12 t 2. A time Series modeling approach, the observation at time t is modeled as a linear combination of previous observations ∗ Examples: P – The autoregressive model: zt = j≥1 πjzt−j + εt – The autoregressive and moving average model: p q P P zt = πjzt−j + θiεt−i + εt j=1 j=1 2 Discounted least squares/general exponential smoothing n X 2 wt[zt − f(t, β)] t=1 • Ordinary least squares: wt = 1. n−t • wt = w , discount factor w determines how fast information from previous observations is discounted. • Single, double and triple exponential smoothing procedures – The constant mean model – The Linear trend model – The quadratic model 3 3.2 Constant Mean Model zt = β + εt • β: a constant mean level 2 • εt: a sequence of uncorrelated errors with constant variance σ . If β, σ are known, the minimum mean square error forecast of a future observation at time n + l, zn+l = β + εn+l is given by zn(l) = β 2 2 2 • E[zn+l − zn(l)] = 0, E[zn+1 − zn(l)] = E[εt ] = σ • 100(1 − λ) percent prediction intervals for a future realization are given by [β − µλ/2σ; β + µλ/2σ] where µλ/2 is the 100(1 − λ/2) percentage point of standard normal distribution.
    [Show full text]
  • Exponential Smoothing
    Key Concepts: Week 3 Lesson 1: Exponential Smoothing Learning Objectives Understand how exponential smoothing treats old and new information differently Understand how changing the alpha or beta smoothing factors influences the forecasts Able to apply the techniques to generate forecasts in spreadsheets Summary of Lesson Exponential smoothing, as opposed to the three other time series models we have discussed (Cumulative, Naïve, and Moving Average), treats data differently depending on its age. The idea is that the value of data degrades over time so that newer observations of demand are weighted more heavily than older observations. The weights decrease exponentially as they age. Exponential models simply blend the value of new and old information. We have students create forecast models in spreadsheets so they understand the mechanics of the model and hopefully develop a sense of how the parameters influence the forecast. The alpha factor (ranging between 0 and 1) determines the weighting for the newest information versus the older information. The “α” value indicates the value of “new” information versus “old” information: As α → 1, the forecast becomes more nervous, volatile and naïve As α → 0, the forecast becomes more calm, staid and cumulative α can range from 0 ≤ α ≤ 1, but in practice, we typically see 0 ≤ α ≤ 0.3 The most basic exponential model, or Simple Exponential model, assumes stationary demand. Holt’s Model is a modified version of exponential smoothing that also accounts for trend in addition to level. A new smoothing parameter, β, is introduced. It operates in the same way as the α. We can also use exponential smoothing to dampen trend models to account for the fact that trends usually do not remain unchanged indefinitely as well as for creating a more stable estimate of the forecast errors.
    [Show full text]
  • Moving Averages and Smoothing Methods QMIS 320 Chapter 4
    DEPARTMENT OF QUANTITATIVE METHODS & INFORMATION SYSTEMS Moving Averages and Smoothing Methods QMIS 320 Chapter 4 Fall 2010 Dr. Mohammad Zainal Moving Averages and Smoothing Methods 2 ¾ This chapter will describe three simple approaches to forecasting a time series: ¾ Naïve ¾ Averaging ¾ Smoothing ¾ Naive methods are used to develop simple models that assume that very recent data provide the best predictors of the future. ¾ Averaging methods generate forecasts based on an average of past observations. ¾ Smoothing methods produce forecasts by averaging past values of a series with a decreasing (exponential) series of weights. QMIS 320, CH 4 by M. Zainal QMIS 320, CH 4 by M. Zainal 1 Moving Averages and Smoothing Methods 3 Forecasting Procedure ¾ You observe Yt at time t ¾ Then use the past data Yt‐1,Yt‐2, Yt‐3 ¾ Obtain forecasts for different periods in the future QMIS 320, CH 4 by M. Zainal Moving Averages and Smoothing Methods 4 Steps for Evaluating Forecasting Methods ¾ Select an appropriate forecasting method based on the nature of the data and experience. ¾ Divide the observations into two parts. ¾ Apply the selected forecasting method to develop fitted model for the first part of the data. ¾ Use the fitted model to forecast the second part of the data. ¾ Calculate the forecasting errors and evaluate using measures of accuracy. ¾ Make your decision to accept or modify or reject the suggested method or develop another method to compare the results. QMIS 320, CH 4 by M. Zainal QMIS 320, CH 4 by M. Zainal 2 Moving Averages and Smoothing Methods 5 Types of Forecasting Methods Three simple approaches to forecasting a time series will be discussed in this chapter: Naïve Methods These methods assume that the most recent information is the best predictors of the future.
    [Show full text]
  • Exponential Smoothing: Seasonality
    CTL.SC1x -Supply Chain & Logistics Fundamentals Exponential Smoothing: Seasonality MIT Center for Transportation & Logistics CTL.SC1x - Supply Chain and Logistics Fundamentals Fundamentals Logistics and Chain - Supply CTL.SC1x Annual Seasonality for Various Products forVarious Seasonality Annual cloth bandages bandages cloth candy chocolate& garden tools garden salads salads Jan Lesson: Exponential Smoothing with Seasonality withSmoothing Seasonality Exponential Lesson: Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Agenda • Double Exponential Smoothing Model n Level & Seasonality • Holt-Winter Model n Level, Trend & Seasonality • Initialization of Parameters • Practical Concerns CTL.SC1x - Supply Chain and Logistics Fundamentals Lesson: Exponential Smoothing with Seasonality 3 Double Exponential Smoothing CTL.SC1x - Supply Chain and Logistics Fundamentals Lesson: Exponential Smoothing with Seasonality 4 Time Series Analysis a • Exponential Smoothing for Level & Seasonality Demand rate time n Double exponential smoothing SF n Introduces multiplicative seasonal term Demand rate Underlying Model: xt = aFt + et time 2 where: et ~ iid (µ=0 , σ =V[e]) Forecasting Model: ˆ ˆ ˆ The most current estimate xt,t+τ = at Ft+τ −P of the appropriate seasonal index Updating Procedure: " x % aˆ = α$ t '+ 1−α aˆ t $ ˆ ' ( )( t−1) F De-seasonalizing F = Multiplicative seasonal index # t−P & t the most recent appropriate for period t ! x $ actual observation. P = Number of time periods within the Fˆ = γ # t &+ 1−γ Fˆ seasonality t # aˆ & ( ) t−P γ = Seasonality
    [Show full text]
  • Fitting and Extending Exponential Smoothing Models with Stan
    Fitting and Extending Exponential Smoothing Models with Stan Slawek Smyl <[email protected]> Qinqin Zhang <[email protected]> International Symposium on Forecasting Riverside, California, USA, 2015 Fitting and Extending Exponential Smoothing Models with Stan Slawek Smyl, Qinqin Zhang Microsoft {slsmyl, qinzh}@microsoft.com Abstract. Time series forecasting is a fundamental arbitrary dynamics and estimating models numerically challenge to various industries like retail and finance. leveraging the power of computation. Most of the classical forecasting models, including regression, ARIMA, exponential smoothing, assume Among various probabilistic programming languages, that the characteristics of input data can be fully Stan (Stan Development Team 2015) is a relatively captured except for mean-zero Gaussian noise terms. mature and comprehensive option. It is a flexible However, this assumption is frequently invalid for programming tool which implements full Bayesian highly volatile data with heavy-tail distributions. statistical inference with Hamiltonian Markov Chain Analytical tractability is a big obstacle combating non- Monte Carlo. In Stan, models are specified in a Gaussian noise terms. But with Stan (http://mc- language with syntax and some conventions similar to stan.org), a probabilistic programming language a mix of R and C++. Then the models are translated to implementing full Bayesian statistical inference, C++ and compiled so that sampling can be executed. implementing innovative time series forecasting algorithms becomes easier. In this study, we discuss a Using Stan as a probabilistic programming tool has few extensions to Holt-Winters Double Exponential several benefits. First, with the interface package Smoothing algorithm including RStan (Stan Development Team 2015), users can easily 1) using a robust error distribution (Student) develop Stan codes combined with R codes using an R- 2) allowing trend and sigma to grow non-linearly with scripting environment.
    [Show full text]
  • Stat Forecasting Methods V6.Pdf
    Driving a new age of connected planning Statistical Forecasting Methods Overview of all Methods from Anaplan Statistical Forecast Model Predictive Analytics 30 Forecast Methods Including: • Simple Linear Regression • Simple Exponential Smoothing • Multiplicative Decomposition • Holt-Winters • Croston’s Intermittent Demand Forecasting Methods Summary Curve Fit Smoothing Capture historical trends and project future trends, Useful in extrapolating values of given non-seasonal cyclical or seasonality factors are not factored in. and trending data. ü Trend analysis ü Stable forecast for slow moving, trend & ü Long term planning non-seasonal demand ü Short term and long term planning Seasonal Smoothing Basic & Intermittent Break down forecast components of baseline, trend Simple techniques useful for specific circumstances or and seasonality. comparing effectiveness of other methods. ü Good forecasts for items with both trend ü Non-stationary, end-of-life, highly and seasonality recurring demand, essential demand ü Short to medium range forecasting products types ü Short to medium range forecasting Forecasting Methods List Curve Fit Smoothing Linear Regression Moving Average Logarithmic Regression Double Moving Average Exponential Regression Single Exponential Smoothing Power Regression Double Exponential Smoothing Triple Exponential Smoothing Holt’s Linear Trend Seasonal Smoothing Basic & Intermittent Additive Decomposition Croston’s Method Multiplicative Decomposition Zero Method Multiplicative Decomposition Logarithmic Naïve Method Multiplicative Decomposition Exponential Prior Year Multiplicative Decomposition Power Manual Input Winter’s Additive Calculated % Over Prior Year Winter’s Multiplicative Linear Approximation Cumulative Marketing End-of-Life New Item Forecast Driver Based Gompertz Method Custom Curve Fit Methods Curve Fit techniques capture historical trends and project future trends. Cyclical or seasonality factors are not factored into these methods.
    [Show full text]
  • Lecture 7: Exponential Smoothing Methods Please Read Chapter 4 and Chapter 2 of MWH Book
    Lecture 7: Exponential Smoothing Methods Please read Chapter 4 and Chapter 2 of MWH Book 1 Big Picture 1. In lecture 6, smoothing (averaging) method is used to estimate the trend-cycle (decomposition) 2. Now, modified smoothing method is used to forecast future values. That means, in general, the averaging is one-sided, as opposed to two-sided 3. Another difference is, we focus on out-of-sample forecasting errors, other than the in-sample fitting errors (residuals) 4. Again, which method works best depends on whether trend or seasonality is present 2 Forecasting Scenario 1. The present is the t-th period 2. Past (in-sample) observations Yt ;Yt−1;:::;Y1 are used to estimate the model 3. Then forecasts Ft+1;Ft+2;::: are computed for the future (out-of-sample) values Yt+1;Yt+2;::: 4. Do not confuse the out-of-sample forecasting errors Yt+1 − Ft+1;Yt+2 − Ft+2;::: with the in-sample fitting errors (residuals) Yt − Ft ;:::;Y1 − F1 3 Averaging Method—Mean This method uses all past data available 1 t Ft+1 = ∑Yi (1) t i=1 1 t+1 Ft+2 = ∑ Yi (2) t + 1 i=1 The mean method works well only when there is no trend or seasonality or change in the mean value. In other words, it assumes the underlying process has a constant mean (stationarity). 4 Averaging Method—Moving Average Moving average, to some degree, can allow for time-varying mean value. It does so by including only the most recent k observations. More explicitly, the MA(k) forecast (not smoother) is 1 t Ft+1 = ∑ Yi (3) k i=t−k+1 1 t+1 Ft+2 = ∑ Yi (4) k i=t−k+2 It follows that 1 F = F + (Y −Y − ) (5) t+2 t+1 k t+1 t k+1 So each new forecast Ft+2 is an adjustment of the immediately preceding forecast Ft+1: The 1 − adjustment term is k (Yt+1 Yt−k+1): Bigger k leads to smaller adjustment (or smoother forecasts).
    [Show full text]